some ideas on quantization

[Hellworm]

I know that x264 has a temporal denoising option, but I just had an idea about making denoising part of the quantisation. If a block has found reference block then the difference is transformed with a 8x8 or 4x4 interger transformation and the result is the quantized, which as a result (should) destroy only useless information like noise. However the quantization process has no knowledge of knowing what in the difference is noise and what isn't, and is therefor independent of the temporal information that the reference frames represent. So I thought it's possible to do a transformation on the blocks that are referenced and analyse how the coefficients change over time and then change the way each coefficient is quantized depending on that information. For example by having a per coefficient deadzone or use the information to help trellis find the right quantisation.

So to the experts, what do you think?

[akupenguin]

Since it's a linear transform, DCT commutes with inter prediction. Thus you don't get any extra information by transforming the reference block.

[Hellworm]

Sorry, I don't understand at all :o, could you explain that more? I didn't meant to transform the reference and compare that to the residual, but to compare the transformation of the original, uncompressend blocks (like it is done in fft3dfilter) and apply that knowledge to the quantization.